No Scale Left Behind: Multi-Scale Autoencoder with Bi-directional Attention for Time Series Anomaly Detection

Jiaheng Guo; Haochen Zhang; Yu-Chao Huang; Jinhao Duan; Nicholas Konz; Tianlong Chen

cs.LGSep 29, 2026

No Scale Left Behind: Multi-Scale Autoencoder with Bi-directional Attention for Time Series Anomaly Detection

Authors: Jiaheng Guo, Haochen Zhang, Yu-Chao Huang, Jinhao Duan, Nicholas Konz, Tianlong Chen

Organizations: UNITES Lab, University of North Carolina at Chapel Hill

Abstract

Time series anomaly detection (TSAD) plays a crucial role in healthcare, finance, industrial monitoring, and other sectors. Within and between these settings, anomalies span vastly different temporal scales, from sub-second point spikes to multi-hour drift patterns. However, most existing TSAD methods commit to a single temporal granularity, and multi-scale designs either analyze different scales in isolation or are constrained to a predefined coarse-to-fine hierarchy, both failing to sufficiently capture multi-scale interactions. To resolve this limitation, we propose Multi-Scale Autoencoder with Cross-Scale Attention for TSAD (MSCAD), a simple yet powerful semi-supervised TSAD framework founded on parallel autoencoder branches corresponding to different patch sizes. A stack of symmetric bidirectional cross-scale attention blocks enables every pair of scales to exchange information before reconstruction without allowing any single scale to be privileged. On the comprehensive TSB-AD benchmark (40 datasets, 530 series), MSCAD achieves large performance gains against 50 baselines across multiple metrics, with VUS-PR of 0.57(+9.6%) on the univariate split and 0.47(+9.3%) on the multivariate split compared to the state-of-the-art.

Figures & tables

Panel of Figure 1 — Figure 1: (Top) Anomaly-event duration distribution on TSB-AD. Short events ( $<16$ timesteps) dominate; long events ( $\geq 128$ timesteps) form a non-trivial $12$ – $14\%$ tail, motivating a multi-scale design. (Bottom) Cross-scale coupling topologies. Existing multi-scale TSAD methods either keep scales independent or impose asymmetric coarse-to-fine flow. MSCAD uses a bidirectional bridge, allowing each scale to exchange information with every other scale symmetrically.

Figure 2: MSCAD architecture. Three parallel patch-Transformer branches at scales $\mathcal{P}=\{4,16,64\}$ (§ 3.3 ) are connected by a stack of $B$ symmetric cross-scale bridges (§ 3.4 ); each branch decodes its bridged latents back to patch space, and the uniform mean of the per-scale reconstruction errors is the window score $f_{\theta}(\mathbf{x})$ .

Figure 3: Per-bucket mean VUS-PR vs. each file’s max anomaly-event duration (10 quantile-spaced buckets) on TSB-AD. Gold shading: MSCAD’s lead over the best single-scale baseline per bucket. Gray band: min–max envelope of single-scale variants

{P}\in\{{4},{16},{64}\}

Configuration	UTS VUS-PR	MTS VUS-PR	Average VUS-PR
No bridge ( $B{=}0$ )	0.5380	0.4412	0.5051
Identity bridge (FFN $+$ residual only)	0.5399	0.4440	0.5073 $\pm$ .0017
Symmetric cross-attention bridge (default)	0.5691 $\pm$ .0018	0.4705 $\pm$ .0019	0.5356 $\pm$ .0018

Table 2: Bridge contribution at the headline configuration (

d{=}256

B{=}2

). The identity variant (FFN+residual but no cross-scale context) is statistically tied with the no-bridge baseline, isolating the cross-scale exchange as the load-bearing component. Entries with

\pm

are

3

-seed mean

\pm

stddev.

Direction	UTS VUS-PR	MTS VUS-PR	Average VUS-PR
c2f (CrossAD-style)	0.5412	0.4529	0.5112
f2c	0.5313	0.4511	0.5041
sym (default)	0.5691	0.4705	0.5356

Table 3: Both directed variants substantially reduce the bridge gain: c2f drops by

0.0244

All from sym , while f2c lands near the no-bridge baseline (

0.5051

in Table 2 ).

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Track	Split	# series	Avg. dim.	Avg. len.	Avg. # ev.	Avg. ev. len.
TSB-AD-U	All	870	1.0	38,814.0	39.7	179.5
	Eval	350	1.0	51,886.7	46.6	321.3
	Tuning	48	1.0	47,143.3	82.6	185.9
TSB-AD-M	All	200	29.3	107,760.5	71.1	582.6
	Eval	180	29.1	108,826.8	67.8	591.3
	Tuning	20	30.4	98,164.1	101.1	504.8

Appendix

Table 4: Aggregate statistics of TSB-AD. The univariate track contains 870 curated series and the multivariate track contains 200 curated series. The headline TSB-AD protocol evaluates on 350 univariate and 180 multivariate held-out series, after selecting hyperparameters on the tuning split.

Source	# series	Eval	Tune	Avg. length	Anom. ratio	Type
UCR ( Wu and Keogh, 2023 )	228	70	6	68,052.9	0.6%	P&Seq
NAB ( Ahmad et al., 2017 )	28	23	5	5,099.8	10.6%	Seq
YAHOO ( Laptev et al., 2015 )	259	30	5	1,561.8	0.6%	P&Seq
IOPS ( IOPS, n.d. )	17	15	2	72,792.3	1.4%	Seq
MGAB ( Thill et al., 2020 )	9	8	1	97,777.8	0.2%	Seq
WSD ( Zhang et al., 2022 )	111	20	5	17,509.5	0.6%	Seq

Appendix

Table 5: TSB-AD-U source breakdown. All time series are univariate. The Eval and Tune columns count files appearing in TSB-AD-U-Eva.csv and TSB-AD-U-Tuning.csv, respectively.

Source	# series	Eval	Tune	Avg. dim.	Avg. length	Anom. ratio	Type
GHL ( Filonov et al., 2016 )	25	23	2	19.0	199,001.0	1.1%	Seq
Daphnet ( Bächlin et al., 2010 )	1	1	0	9.0	38,774.0	5.9%	Seq
Exathlon ( Jacob et al., 2021 )	27	25	2	20.5	60,878.4	9.8%	Seq
Genesis ( von Birgelen and Niggemann, 2018 )	1	1	0	18.0	16,220.0	0.3%	Seq
OPP ( Roggen et al., 2010 )	8	7	1	248.0	17,426.8	4.1%	Seq
SMD ( Su et al., 2019 )	22	20	2	38.0	25,466.4	3.8%	Seq

Appendix

Table 6: TSB-AD-M source breakdown. The Avg. dim. column reports the average number of observed channels, excluding the binary label column.

Group	Hyperparameter	Value
Architecture	Window length $W$	$128$
	Sliding stride $\tau$	$P_{\min}/2=2$
	Patch sizes $\mathcal{P}$	$\{4,16,64\}$
	Number of scales $S$	$3$
	Token dim $d$	$256$
	Attention heads $h$	$4$

Appendix

Table 7: Reference hyperparameters for the headline MSCAD (symmetric cross-attention,

d{=}256

B{=}2

). Sensitivity around these values is reported in Table 13 and Appendix F .

Univariate split (TSB-AD-U, $23$ datasets, $350$ series, $39$ baselines)
Method	VUS-PR	VUS-ROC	Range-F1	AUC-PR	AUC-ROC	Point-F1
Sub-PCA	0.42	0.76	0.41	0.37	0.71	0.42
KShapeAD	0.40	0.76	0.40	0.35	0.74	0.39
POLY	0.39	0.76	0.35	0.31	0.73	0.37
Series2Graph	0.39	0.80	0.35	0.33	0.76	0.38
MOMENT (FT)	0.39	0.76	0.35	0.30	0.69	0.35

Appendix

Table 8: Full baseline comparison on TSB-AD-U against MSCAD. Bold marks the best per column.

Multivariate split (TSB-AD-M, $17$ datasets, $180$ series, $33$ baselines)
Method	VUS-PR	VUS-ROC	Range-F1	AUC-PR	AUC-ROC	Point-F1
CNN	0.31	0.76	0.37	0.32	0.73	0.37
OmniAnomaly	0.31	0.69	0.37	0.27	0.65	0.32
PCA	0.31	0.74	0.29	0.31	0.70	0.37
LSTMAD	0.31	0.74	0.38	0.31	0.70	0.36
USAD	0.30	0.68	0.37	0.26	0.64	0.31

Appendix

Table 9: Full baseline comparison on TSB-AD-M against MSCAD. Bold marks the best per column.

Univariate split ( TSB-AD-U , $23$ datasets, $350$ series)
Dataset	$n$	VUS-PR	VUS-ROC	Range-F1	AUC-PR	AUC-ROC	Point-F1
CATSv2	1	0.3414	0.7490	0.3135	0.5428	0.7533	0.6422
Daphnet	1	0.5166	0.9489	0.5145	0.5253	0.9521	0.5650
Exathlon	30	0.9531	0.9975	0.9372	0.9516	0.9975	0.9687
IOPS	15	0.2950	0.8927	0.3269	0.3732	0.8953	0.4213
LTDB	8	0.6291	0.7758	0.6185	0.5633	0.7638	0.6083

Appendix

Table 10: Per-dataset breakdown of MSCAD on TSB-AD, averaged over seeds

\{2026,2027,2028\}

. Rows are datasets within each split; columns match the six metrics from the main results table.

Metric	UTS ( $n{=}350$ )	MTS ( $n{=}180$ )	Overall ( $n{=}530$ )
AUC-PR	$0.5366\pm 0.0017$	$0.4470\pm 0.0015$	$0.5062\pm 0.0016$
AUC-ROC	$0.8911\pm 0.0003$	$0.7835\pm 0.0016$	$0.8545\pm 0.0003$
VUS-PR	$0.5691\pm 0.0018$	$0.4705\pm 0.0019$	$0.5356\pm 0.0018$
VUS-ROC	$0.9030\pm 0.0002$	$0.7910\pm 0.0015$	$0.8649\pm 0.0005$
Point-F1	$0.5682\pm 0.0020$	$0.4878\pm 0.0024$	$0.5409\pm 0.0021$
PA-F1	$0.6559\pm 0.0021$	$0.5421\pm 0.0040$	$0.6173\pm 0.0025$

Appendix

Table 11: MSCAD headline configuration,

3

-seed mean

\pm

standard deviation across seeds

\{2026,2027,2028\}

on every metric reported by TSB-AD’s get_metrics routine. Per-seed standard deviation is

\leq 0.0040

on every cell.

Method	Params	UTS time/series	MTS time/series	Total per seed
MSCAD attn ( $d{=}256$ , depth $2$ )	5.97 M	14.0 s	7.8 min	24.8 h
MSCAD attn ( $d{=}128$ , depth $1$ )	1.01 M	11.6 s	4.4 min	14.4 h
MSCAD no-bridge ( $d{=}256$ , depth $2$ )	5.18 M	8.3 s	3.5 min	11.2 h

Appendix

Table 12: Compute and parameter cost. Wall times are per-series end-to-end (training + inference) on a single NVIDIA RTX PRO 6000 Blackwell GPU; per-seed totals are summed across the full TSB-AD evaluation split (350 UTS + 180 MTS). MSCAD variants share the reference configuration (

\text{win\_size}{=}128

\text{patch\_sizes}{=}[4,16,64]

4

heads).

Figure 4: MSCAD anomaly-score visualisations on representative TSB-AD series. For each panel, the time series is plotted on top with the labelled anomaly window shaded; the bottom trace is MSCAD’s raw per-timestep score. The score concentrates inside the labelled regions on both short point anomalies and long contextual anomalies.

Figure 5: Per-scale reconstructions on a representative TSB-AD-U series. Left to right: the original window, the reconstruction at $P{=}4$ , $P{=}16$ , $P{=}64$ , and the per-patch squared error at each scale. The complementary error profiles across scales motivate the uniform fusion used by MSCAD.

	$d_{\rm model}$ ( $B{=}2$ , $W{=}128$ )				Bridge depth $B$ ( $d{=}256$ , $W{=}128$ )					Window $W$ ( $d{=}256$ , $B{=}2$ )
Metric	64	128	256	512	0	1	2	3	4	64	128	256
UTS	0.54	0.56	0.57	0.57	0.55	0.56	0.57	0.56	0.56	0.54	0.57	0.53
MTS	0.45	0.46	0.47	0.46	0.45	0.47	0.47	0.47	0.46	0.46	0.47	0.42
All	0.51	0.53	0.54	0.53	0.52	0.53	0.54	0.53	0.53	0.51	0.54	0.49

Appendix

Table 13: Hyperparameter sensitivity around the headline (

d{=}256

B{=}2

W{=}128

, attention bridge). Each block sweeps one HP holding the rest at its block-default. Bold marks the default in each block. Capacity saturates at

d{=}256

, depth saturates at

B{=}2

, and window is inverted-U at

W{=}128

Scale set	UTS VUS-PR	MTS VUS-PR	Average VUS-PR
$\|\mathcal{P}\|{=}2$ , $\mathcal{P}{=}\{4,64\}$	0.5599	0.4557	0.5245
$\|\mathcal{P}\|{=}3$ , $\mathcal{P}{=}\{4,16,64\}$ (default)	0.5691	0.4705	0.5356
$\|\mathcal{P}\|{=}4$ , $\mathcal{P}{=}\{4,16,32,64\}$	0.5657	0.4698	0.5331

Appendix

Table 14: Scale-set composition. Effect of varying the number of patch scales

|\mathcal{P}|

at the headline configuration (

d{=}256

B{=}2

). The two-scale variant uses

\{4,64\}

, the three-scale default uses

\{4,16,64\}

, and the four-scale variant uses

\{4,16,32,64\}

. The default is best on All VUS-PR; dropping the middle scale loses MTS coverage, and adding a fourth scale adds parameters without gain.

Gate variant	UTS VUS-PR	MTS VUS-PR	All VUS-PR
Uniform $1/S$ (default)	0.5691	0.4705	0.5356
Learned MLP, all four features	0.5636	0.4782	0.5346
Learned MLP, random features	0.5636	0.4783	0.5347
Learned MLP, time-only $\{\mu,\sigma\}$	0.5647	0.4738	0.5338
Learned MLP, freq-only $\{f_{\mathrm{dom}},h_{\mathrm{spec}}\}$	0.5630	0.4750	0.5331
Learned MLP, drop $\mu$	0.5622	0.4707	0.5312

Appendix

Table 15: Fusion-gate ablation. VUS-PR on TSB-AD with the uniform

1/S

default replaced by learned input-adaptive gates parameterized over four per-window statistics: window mean

\mu

, standard deviation

\sigma

, dominant-frequency

f_{\mathrm{dom}}

, and spectral entropy

h_{\mathrm{spec}}

. “Random features” substitutes shuffled values for the four statistics; “time-only” and “freq-only” restrict the input to

\{\mu,\sigma\}

and

\{f_{\mathrm{dom}},h_{\mathrm{spec}}\}

respectively; “drop

X

” removes one feature. Reported on the headline configuration.

Bridge operator	UTS VUS-PR	MTS VUS-PR	All VUS-PR
No bridge ( $B{=}0$ )	0.5380	0.4412	0.5051
Identity (FFN $+$ residual only)	0.5399 $\pm$ .0004	0.4440 $\pm$ .0043	0.5073 $\pm$ .0017
Cross-attention	0.5691 $\pm$ .0018	0.4705 $\pm$ .0019	0.5356 $\pm$ .0018
Mean-pool	0.5654 $\pm$ .0010	0.4682 $\pm$ .0042	0.5324 $\pm$ .0014
Sum-pool	0.5702 $\pm$ .0022	0.4513 $\pm$ .0311	0.5312 $\pm$ .0074
Max-pool	0.5677 $\pm$ .0009	0.4630 $\pm$ .0040	0.5321 $\pm$ .0017

Appendix

Table 16: Bridge operator. Effect of replacing the cross-attention operator inside the symmetric bridge with simpler aggregators, holding the FFN+residual wrapper and the symmetric concatenation context fixed (

d{=}256

B{=}2

). All non-trivial operators beat the no-bridge / identity baselines by a similar margin, supporting the § 4.3 claim that the cross-scale exchange is the load-bearing piece. Entries with

\pm

are 3-seed mean

\pm

standard deviation.

$\lambda_{\mathrm{div}}$	UTS VUS-PR	MTS VUS-PR	All VUS-PR
$0$ (default)	0.5691	0.4705	0.5356
$0.1$	0.5388	0.4559	0.5207

Appendix

Table 17: Diversity-loss sweep. Effect of adding a per-scale diversity regularizer with weight

\lambda_{\mathrm{div}}

to the reconstruction loss. The default

\lambda_{\mathrm{div}}{=}0

outperforms the regularized variant on every metric, supporting the § 3.5 design choice to omit auxiliary loss terms. Reported on the symmetric cross-attention bridge with uniform fusion, seed

2026

Explore similar work

PaAno+: Multiscale Encoding and Cross-Variable Attention for Time Series Anomaly Detection
Jun 18, 2026Youji Zhu, Hongbing Wang, Wenchao Liu +2Time-Series Anomaly Detection Time Series
School of Mathematical Sciences, Guizhou Normal University, Guiyang 550025, China. · School of Big Data and Computer Science, Guizhou Normal University, Guiyang 550025, China.
VAN-AD: Visual Masked Autoencoder with Normalizing Flow For Time Series Anomaly Detection
Mar 27, 2026PengYu Chen, Shang Wan, Xiaohou Shi +3Time-Series Anomaly Detection Unsupervised Detection
School of Computer Science (National Pilot Software Engineering School), Beijing University of Posts and Telecommunications, Beijing 100876, China · China Telecom Research Institute Beijing, China · Department of Computer Science, Missouri University of Science and Technology, Rolla, MO 65409 USA
LEFT: Learnable Fusion of Tri-view Tokens for Unsupervised Time Series Anomaly Detection
Feb 9, 2026Dezheng Wang, Tong Chen, Guansong Pang +3Time-Series Anomaly Detection Cross-View Fusion
School of Automation, Southeast University Nanjing, China · School of Electrical Engineering and Computer Science, The University of Queensland Brisbane, Australia · School of Computing and Information Systems, Singapore Management University Singapore

Jun 18, 2026, cs.LGJ/K move · Enter open · S save

PaAno+: Multiscale Encoding and Cross-Variable Attention for Time Series Anomaly Detection

Youji Zhu, Hongbing Wang, Wenchao Liu, Xiaodong Liu+1

School of Mathematical Sciences, Guizhou Normal University, Guiyang 550025, China. · School of Big Data and Computer Science, Guizhou Normal University, Guiyang 550025, China.

Time-Series Anomaly Detection Time Series Progressive Attention Fusion Multiscale Cross-Attention Representation Learning

Open paper AskPDF arXiv

Time-series anomaly detection has significant practical value for industrial and medical monitoring, as well as other critical domains. Current Transformer- and large-model-based detection approaches incur excessive computational overhead, while existing lightweight alternatives are constrained by insufficient feature extraction and inadequate modeling of dependencies across multivariate variables. To mitigate the above drawbacks, this study develops a lightweight, efficient anomaly detection model, dubbed PaAno, within the patch-oriented representation learning paradigm. In the encoder module, a multiscale feature-extraction backbone is constructed using convolutional kernels with differentiated receptive fields to capture hierarchical temporal characteristics; subsequent cross-scale adaptive attention aggregation, combined with residual connection optimization, further stabilizes feature representation learning. A cross-variable fusion attention module is embedded to explicitly characterize inter-variable correlations, empowering the model to identify anomalous patterns amid intricate operational conditions. Moreover, a novel pretext task based on temporal patch-window sorting is customized to uncover intrinsic structural properties of time series, and triplet loss is leveraged to optimize the patch embedding space for enhanced feature discrimination. Extensive experiments on the TSB-AD benchmark demonstrate that the proposed PaAno achieves state-of-the-art detection accuracy on both univariate and multivariate tasks, yielding significant performance gains across evaluation metrics, including VUS-PR, relative to the original PaAno. Leveraging a compact network design, the presented model achieves favorable computational efficiency, enabling deployment on resource-limited terminals for real-time anomaly inference.

Abstract

Explore similar work

PaAno+: Multiscale Encoding and Cross-Variable Attention for Time Series Anomaly Detection

VAN-AD: Visual Masked Autoencoder with Normalizing Flow For Time Series Anomaly Detection

LEFT: Learnable Fusion of Tri-view Tokens for Unsupervised Time Series Anomaly Detection